Traffic Speed Forecasting using STGCN

Predicting urban traffic speeds across 207 road sensors using Spatio-Temporal Graph Convolutional Networks.

---

Overview

This project implements a Spatio-Temporal Graph Convolutional Network (STGCN) to forecast future traffic speeds using historical sensor data from the METR-LA dataset.

The model jointly captures:

• Spatial dependencies between road segments via graph convolution
• Temporal patterns across time steps via temporal convolution

---

Problem Statement

Predict traffic speed for the next 30 minutes based on past observations from 207 road sensors on the Los Angeles highway network (METR-LA dataset).

---

Model Architecture

Input Sequence (T timesteps × 207 nodes)
        │
  ┌─────▼─────┐
  │ STGCN Block 1 │  ← Graph Conv (spatial) + Temporal Conv
  └─────┬─────┘
        │
  ┌─────▼─────┐
  │ STGCN Block 2 │  ← Graph Conv (spatial) + Temporal Conv
  └─────┬─────┘
        │
  ┌─────▼──────────┐
  │ Fully Connected │  ← Output layer
  └─────┬──────────┘
        │
  Predicted Speed (next 30 min)

Each STGCN block combines:

• Graph Convolution — propagates features across the road network adjacency graph
• Temporal Convolution — extracts time-series patterns with gated activation

---

Tech Stack

Component	Tool
Deep Learning	PyTorch
Data Processing	NumPy, Pandas
API Backend	FastAPI
Frontend UI	Streamlit
Containerization	Docker (optional)

---

Project Structure

STGCN_FINAL/
├── api/                # FastAPI backend (inference endpoint)
├── app/                # Streamlit UI (interactive dashboard)
├── notebooks/          # Experiments and training notebooks
├── assets/             # Screenshots and demo media
└── README.md

---

Getting Started

1. Install Dependencies

pip install -r requirements.txt

2. Start the API

uvicorn api.api:app --reload

3. Launch the Streamlit App

streamlit run app/app.py

---

Demo

The Streamlit dashboard includes an interactive slider that lets you scrub through X_train sequences in real time — demonstrating the model's ability to handle dynamic, changing inputs and produce live speed forecasts across the sensor network.

---

Results

• Model successfully learns joint spatio-temporal dependencies across the road graph
• Produces smooth, realistic traffic speed forecasts over a 30-minute horizon
• Interactive demo validates generalization across varied input windows

---

Notes

• Dataset not included due to size constraints — download METR-LA from the official source
• Model was trained and evaluated on the METR-LA benchmark
• Intractive slider to show live data using X_test.npm to show that model can also perform on dynamic data in streamlite
• you have to also set X_test.npm in your artifacts folder so that while runing streamlit doesnot give error

---

Future Improvements

• Integrate a real-time data ingestion pipeline
• Hyperparameter tuning for improved MAE/RMSE
• Live cloud deployment (AWS / GCP)
• Extend to multi-step forecasting horizons (60 min, 120 min)

---

License

This project is for academic purposes only.